分布式用户隐私保护可调节的云服务个性化QoS预测模型

Distributed user privacy preserving adjustable personalized QoS prediction model for cloud services

个性化服务质量(Qo S,quality of service)预测是构建高质量云服务系统的重要环节,传统基于协同过滤方法采用集中式的训练模式难以保护用户隐私,为了在获取高准确预测效果的同时有效保护用户隐私,提出分布式用户隐私保护可调节的云服务个性化QoS预测模型(DUPPA)。该模型采用“服务器-多用户”架构,服务器协调多个用户,处理多用户上传模型梯度和下载全局模型的请求并维护全局模型参数。为进一步保护用户隐私,提出用户隐私程度调节策略,通过调节本地模型参数初始化比例、梯度上传比例以平衡隐私程度和预测精度。在本地模型初始化阶段,用户计算本地模型与全局模型的差值矩阵,并选择差值矩阵中数值较大元素所对应的全局模型参数初始化本地模型参数;在梯度上传阶段,用户可选择部分重要的梯度上传至服务器来满足不同应用场景对隐私保护的需求。为了评估DUPPA的隐私程度,提出针对分布式矩阵分解模型梯度共享方案的数据重构攻击方法。实验结果表明,当DUPPA在梯度上传比例为0.1、本地模型参数初始化比例为0.5时,预测的平均绝对误差(MAE,mean absolute error)和均方根误差(RMSE,root mean square error)较传统的集中式矩阵分解模型分别降低了1.20%和0.91%;当DUPPA的梯度上传比例为0.1时,隐私程度至少是梯度上传比例为1时的5倍;当DUPPA的本地模型参数初始化比例为0.5时,隐私程度至少是本地模型参数初始化比例为1时的3.44倍。

Personalized quality of service(QoS)prediction is crucial for developing high-quality cloud service system.However,the traditional collaborative filtering method,based on centralized training,presents challenges in protecting user privacy.In order to effectively protect user privacy while obtaining highly accurate prediction effect,a distributed user privacy adjustable personalized QoS prediction model for cloud services(DUPPA)was proposed.The model adopted a“server-multi-user”architecture,in which the server coordinated multiple users,handled multiple users’requests for uploading model gradients and downloading global model,and maintained global model parameters.To further protect user privacy,a user privacy adjustment strategy was proposed to balance privacy and prediction accuracy by adjusting the initialization proportion of local model parameters and gradient upload proportion.In the local model initialization stage,the user calculated the difference matrix between the local model and the global model,and selected the global model parameters corresponding to the larger elements in the difference matrix to initialize the local model parameters.In the gradient upload stage,the user can select some important gradients to upload to the server to meet the privacy protection requirements of different application scenarios.To evaluate the privacy degree of DUPPA,a data reconstruction attack method was proposed for the distributed matrix factorization model gradient sharing scheme.The experimental results show that when DUPPA sets the gradient upload proportion to 0.1 and the local model parameter initialization proportion to 0.5,the predicted MAE and RMSE are reduced by 1.27%and 0.91%,respectively,compared with the traditional centralized matrix factorization model.Besides,when DUPPA sets the gradient upload proportion to 0.1,the privacy degree is 5 times higher than when the gradient upload proportion is 1.And when DUPPA sets the local model parameter initialization proportion to 0.5,the privacy degree is 3.44 times higher than when the local model parameter initialization proportion is 1.